ZixunWang/rankseg-benchmark-models

RankSEG Benchmark Models

Segmentation models whose test-set probabilities are hosted in ZixunWang/rankseg-benchmark and used by the Quick suite of rankseg-benchmark. They are the models used in the RankSEG-RMA (NeurIPS 2025) and RankSEG-DEP (NeurIPS 2026) papers.

Each directory mirrors the dataset layout and contains:

• model.safetensors: fp32 weights only (optimizer and scheduler states removed);
• config.json: model, data, loss, schedule, optimizer and inference settings, the training epoch, and the SHA-256 of the original training checkpoint.

Directory	Model	Classes	Test inputs	Argmax Dice / IoU (%)
pascal_voc/	UPerNet, ConvNeXt-B (IN-22k → IN-1k, 384)	21	1,449 val images	91.03 / 87.82
cityscapes/	UPerNet, ConvNeXt-B (IN-22k → IN-1k, 384)	19	500 val images	82.61 / 75.66
ade20k/	UPerNet, ConvNeXt-B (IN-22k → IN-1k, 384)	150	2,000 val images	63.98 / 56.94
deepglobe_land/	UPerNet, ConvNeXt-B (IN-22k → IN-1k, 384)	6 (+ unknown ignored)	163 test images	68.60 / 59.49
kits/fold{0-4}/	DeepLabV3+, ResNet-50d	2	5-fold CV, 2D slices	61.16 / 54.18
lits/fold{0-4}/	DeepLabV3+, ResNet-50d	2	5-fold CV, 2D slices	47.38 / 38.34

Scores are those of the Quick benchmark on the hosted probabilities: natural images average per-class scores over the classes present in each image, then over images; KiTS and LiTS average foreground slice scores within each case, then over cases and folds.

Loading

modeling.py is a self-contained definition of both architectures (requires torch, timm, safetensors and huggingface_hub; tested with torch 2.4 and timm 1.0.14). It reproduces the original training code's outputs exactly.

from huggingface_hub import hf_hub_download
import importlib.util

path = hf_hub_download("ZixunWang/rankseg-benchmark-models", "modeling.py")
spec = importlib.util.spec_from_file_location("modeling", path)
modeling = importlib.util.module_from_spec(spec)
spec.loader.exec_module(modeling)

model, config = modeling.from_pretrained("lits/fold0", device="cuda")
probs = model.predict(image)  # preprocessed (B, C, H, W) -> softmax probabilities

model(x) returns logits for an input of exactly config["data"]["crop_size"]; model.predict applies the sliding-window inference used for the hosted probabilities.

Preprocessing and inference

• Natural images (VOC, Cityscapes, ADE20K, DeepGlobe): RGB padded to at least data.crop_size, normalized with mean [123.68, 116.28, 103.53] and std [58.40, 57.12, 57.38] (0–255 scale); single-scale sliding-window inference with window data.crop_size and stride crop_size × 2/3, no flipping (config["inference"]).
• CT slices (KiTS, LiTS): one input channel, center-cropped or padded (with window_low) to 384 × 384, Hounsfield units clipped to [data.window_low, data.window_high] and mapped linearly to [−1, 1]. KiTS: window [−200, 300]; LiTS: window [−60, 140].
• Sliding-window probabilities as used for the hosted data: model.predict(image).
• Fold k of KiTS and LiTS is evaluated on the cases held out in fold k (data.fold, data.num_cases), matching kits/fold{k} and lits/fold{k} in the probability dataset.

License

The weights and config.json files are released under CC BY-NC-SA 4.0 for non-commercial research use only; modeling.py is released under the MIT License. The weights are provided as is, without warranty, and are not intended for clinical use.

Each model was trained on a third-party dataset, and you are responsible for complying with that dataset's own terms. Some of these terms are stricter than CC BY-NC-SA 4.0. For example, Cityscapes may be used only for non-commercial purposes and may not be redistributed.

Training data	Source and terms
PASCAL VOC 2012	host.robots.ox.ac.uk/pascal/VOC
Cityscapes	cityscapes-dataset.com
ADE20K	ADE20K terms of use
DeepGlobe Land Cover	DeepGlobe 2018 challenge
KiTS19	kits19 (CC BY-NC-SA 4.0)
LiTS	LiTS challenge (CC BY-NC-SA 4.0)
Backbones	ImageNet-pretrained convnext_base.fb_in22k_ft_in1k_384 and resnet50d from timm

Citation

@inproceedings{wang2025rankseg,
  title={RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation},
  author={Wang, Zixun and Dai, Ben},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}

@inproceedings{wang2026rankseg,
  title={On the Relaxation of Conditional Independence Assumption for Image Segmentation},
  author={Wang, Zixun and Dai, Ben},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}